融合信息增益与和声搜索的语音情感特征选择
发布时间：2018-09-19 14:39
【摘要】：语音特征集通常具有较高的维度,高维度的特征集不仅包含噪声数据和冗余数据,影响情感识别精度,而且在分类识别的过程中将花费大量的计算开销.如何从较高维度的特征集中选择出规模更小,性能较优的特征子集对语音情感识别系统具有重要作用.本文融合过滤式和封装式两种筛选策略,提出信息增益与和声搜索算法相结合的方法进行语音情感特征选择.试验结果表明,采用过滤和封装相结合的两步策略进行特征选择,综合了过滤策略的低时间开销和封装策略的高识别率的优点,而且可以选择出较原始数据集维度更低且分类性能较好的特征子集.
[Abstract]:Speech feature sets usually have a higher dimension. The high dimension feature sets not only contain noise data and redundant data which affect the accuracy of emotion recognition but also spend a lot of computational overhead in the process of classification and recognition. How to select a feature subset with smaller scale and better performance from a higher dimension feature set plays an important role in speech emotion recognition system. In this paper, two filtering strategies, filtering and encapsulation, are combined, and a method combining information gain and harmonic search algorithm is proposed to select speech emotion features. The experimental results show that the advantages of low time cost of filtering strategy and high recognition rate of packaging strategy are combined with the two-step strategy of filtering and encapsulation for feature selection. Moreover, we can select a feature subset with lower dimension and better classification performance than the original data set.
【作者单位】： 合肥工业大学计算机与信息学院福祉科技实验室;合肥工业大学工业与装备技术研究院;安徽建筑大学电子与信息工程学院;
【基金】：国家自然科学基金面上项目(61370219)资助 国家高等学校学科创新引智计划项目(B14025)资助 情感计算与先进智能机器安徽省重点实验室开放课题项目(ACAIM160103)资助
【分类号】：TN912.34
本文编号：2250424
[Abstract]:Speech feature sets usually have a higher dimension. The high dimension feature sets not only contain noise data and redundant data which affect the accuracy of emotion recognition but also spend a lot of computational overhead in the process of classification and recognition. How to select a feature subset with smaller scale and better performance from a higher dimension feature set plays an important role in speech emotion recognition system. In this paper, two filtering strategies, filtering and encapsulation, are combined, and a method combining information gain and harmonic search algorithm is proposed to select speech emotion features. The experimental results show that the advantages of low time cost of filtering strategy and high recognition rate of packaging strategy are combined with the two-step strategy of filtering and encapsulation for feature selection. Moreover, we can select a feature subset with lower dimension and better classification performance than the original data set.
【作者单位】： 合肥工业大学计算机与信息学院福祉科技实验室;合肥工业大学工业与装备技术研究院;安徽建筑大学电子与信息工程学院;
【基金】：国家自然科学基金面上项目(61370219)资助 国家高等学校学科创新引智计划项目(B14025)资助 情感计算与先进智能机器安徽省重点实验室开放课题项目(ACAIM160103)资助
【分类号】：TN912.34
【相似文献】
相关期刊论文 前1条
1 张小康;帅建梅;史林;;基于加权信息增益的恶意代码检测方法[J];计算机工程;2010年06期
，本文编号：2250424
本文链接：https://www.wllwen.com/kejilunwen/xinxigongchenglunwen/2250424.html
